The weights a culinary recommender learned say more about diners than about restaurants.
Every recommendation screen hides a set of numbers that decide what a person sees. A 2015 paper on a mobile culinary recommendation system published those numbers, and they are worth reading as a statement about how people actually choose a meal.
The system compares three criteria: distance, price, and rating. After the pairwise comparisons were processed, the resulting eigenvector gave each criterion the following share of influence.
The ranking is blunt. Reputation, the signal most food applications put at the centre of the screen, carries the least weight of the three. Two practical constraints, what a meal costs and how far it is, carry the rest. That matches how the input flow is built: the user supplies a budget and a travel radius first, and the candidate list is trimmed before quality is even considered.
Distance here comes from the Haversine formula, which measures the straight line between two points on a sphere. It ignores roads, one way streets, and rivers. For a five kilometre radius, that approximation is acceptable, but it quietly assumes that a place is equally reachable in every direction.
The weights come from comparisons made by the people who configured the system, not from observing a population of diners. A budget conscious student and a diner entertaining guests would not produce the same matrix. The useful lesson for anyone building a recommender is structural rather than numeric: keep the weights as an explicit input that can differ per user, because a single fixed matrix encodes one person's priorities for everyone.
Pinandito et al. (2015). Alternatives Weighting in Analytic Hierarchy Process of Mobile Culinary Recommendation System Using Fuzzy. ARPN Journal of Engineering and Applied Sciences.
